Program

• Paper 663:
  Yuantao Fan, Nuwan Gunasekara, Yibin Sun, Aurora Esteban Toscano, Zhenkan Wang, Slawomir Nowaczyk
  "Online Continual Learning with Domain-Balanced Replay for Energy Forecasting in Heavy-Duty Electric Vehicles"

• Paper 409
  Diogo Risca, Afonso Lourenço, Ricardo Martins, Goreti Marreiros
  "Collaborative Streaming Anomaly Detection with Interactive Explanations and Ensemble Consensus"

• Paper 399
  Vítor Crista, Afonso Lourenço, Diogo Martinho, Goreti Marreiros
  "Streaming Hierarchical Inference with Tabular Foundation Models"

• Paper 200
  Sjoerd van Straten, Marwan Hassani
  "Efficient Online Continual Foundation Model Fine-Tuning for Predictive Process Monitoring"

• Paper 280
  Malavika Suresh, Ikechukwu Nkisi-Orji, Nirmalie Wiratu
  "Relative Parameter Importance in Task-Agnostic Replay-Free Continual Learning"

• Paper 405
  Chiara Thien Thao Nguyen Ba, Lorenzo Iovine, Giacomo Ziffer, Emanuele Della Valle
  "Adaptive Satellite Image Analysis under Temporal and Spatio-Temporal Domain Shifts"

• Paper 419
  Tobias Callies, Eirini Ntoutsi, Arthur Zimek
  "Inference-Time Prior Adaptation for Imbalanced Data Streams under Concept Drift"

• Paper 660
  Helena Moutinho, Rita P. Ribeiro, João Gama
  "Growing Smarter, Not Larger: Instance Selection for Streaming Regression Rule Learners"

• Paper 661
  Francesco Panetta, Federico Di Valerio, Michela Proietti
  "Saliency-Guided Replay for Continual Learning in Deep Spiking Neural Networks"

Abstract:  Streaming machine learning and continual learning have developed rich bodies of literature, active research communities, wide interest from industrial practitioners, and increasingly important roles in modern AI. Both fields address a fundamental challenge: how can learning systems adapt (quickly) in the context of dynamic data, without losing previously acquired knowledge? Pushed by their modern relevance, these areas are merging. However, in combining ideas, we inherit assumptions and unresolved issues from past research endeavours, while leaving some conceptual and methodological gaps that are arguably hindering the efficacy of solutions produced by the community.

This talk takes a fresh look at streaming continual learning. We revisit terminology and foundational assumptions, critically examine progress to date, and identify outstanding challenges and open questions. We then explore the opportunities and implications of studying these problems in a modern context, including a discussion of foundation models and agentic learning (such as reinforcement learning). We conclude by outlining a number of current directions, with the intention to encourage progress in streaming continual learning in a way that ensures continued relevance, as the field and its community move forward.

Bio: Jesse Read is a Professor in the Computer Science Laboratory (LIX) of Ecole Polytechnique in France. He obtained his PhD from the University of Waikato (New Zealand) in 2010, followed by postdoctoral research at Carlos III University (Spain) and Aalto University (Finland), before finally settling in France in 2016. His research interests include learning from evolving data streams, including in the context of data stream learning, reinforcement learning and autonomous agents; probabilistic machine learning; and multi-output models -- among several industrial applications such as predictive maintenance and models of complex energy systems; and applications in medicine and the natural sciences. He has co-authored over 100 scientific publications and has been involved in developing several open-source software projects.

Website
GoogleScholar
Scopus